Firecrawl

What Firecrawl is

Firecrawl is plumbing, not an app. You send it a URL and it sends back the page as markdown, HTML, a screenshot or JSON matching a schema you supply. Send it a domain instead and Crawl walks the whole site, or Map enumerates the URLs. Around that sit Search (web search with optional full-page content), Interact (a browser sandbox billed by the minute), Monitor (re-check pages on a schedule), Batch Scrape for thousands of URLs at once, and Agent, which is still in preview. Firecrawl's own tagline calls it a context API for searching, scraping and interacting with the web at scale, and that is a fair description.

There are official SDKs for Python, Node.js, Go, Rust, Java and Elixir, a CLI, and an MCP server that the vendor's homepage and docs name for Cursor, Claude Code, Codex and Windsurf. There is no workflow builder and no visual scraper. You get an API key and you write code.

The company runs two things that share a name. The GitHub repository is AGPL-3.0, had roughly 176,000 stars when we checked in September 2026 and had been pushed to the same day, so the open-source side is real and actively maintained. It is not, however, the same product as the cloud. Firecrawl's own self-hosting guide says the default stack ships bundled Playwright and fetch processing with a PostgreSQL queue, and that the Fire-engine anti-bot service, screenshots, page actions, Agent, Interact and the specialised output formats are not in the box. Open source here means you can run a crawler. It does not mean you can run Firecrawl.

What we checked was the vendor's pricing page, its machine-readable pricing file, the billing documentation, the rate-limit tables, the self-hosting guide, the benchmarks page and the repository. We did not run the API, so nothing below is a measured latency or a success rate we observed. The coverage and speed claims on the homepage are the vendor's, and we report them as such. What we can check without an account is the money, and the credit model is the part buyers get wrong most often.

Pricing

Firecrawl publishes its prices in full, including a machine-readable pricing file at firecrawl.dev/pricing.md, which is more than most of this category manages. Five self-serve tiers and one quote-only tier, all priced per API key rather than per seat. One caveat on reading the page itself: the rendered pricing page showed us only the annual-billed figures ($16, $83, $333 and $599 a month, each labelled billed yearly), with month-to-month prices behind a toggle. The monthly figures below come from the vendor's own pricing.md, and they reconcile exactly with the "Save $38 / $198 / $798 / $1,798" labels on the rendered page, so we are confident in them. Just do not read $16 as the monthly price.

All figures in US dollars, read from firecrawl.dev/pricing, firecrawl.dev/pricing.md and docs.firecrawl.dev/billing in September 2026. Support is listed as basic on Hobby, standard on Standard, priority on Growth and Scale, and dedicated with an SLA on Enterprise; the Free tier lists no support line. Scale checks out self-serve without a sales call. Enterprise adds zero-data retention, SSO and bulk discounts.

How to budget credits correctly

The plan prices are the easy part. Firecrawl bills in credits, and the headline rate of 1 credit a page is the floor rather than the norm. Two things follow from that. First, the price per credit falls steeply as you climb the ladder, and top-up credits always cost more than the plan credits they sit on. Second, the surcharges for the features most people buy this tool for change the arithmetic by a factor of five. The two tables below show both.

The Hobby-to-Standard step is the one to notice: plan credits get nearly four times cheaper ($0.0038 to $0.00099) for a bill that is 5.2 times larger, and there is no rung between 5,000 and 100,000 credits. Someone who needs 20,000 credits a month can sit on Hobby and buy 15 top-ups for $75, paying $94 in total, or move to Standard at $99 for 100,000 credits they will mostly not use. The second is better value per credit for roughly the same money, which is presumably the point, but it is a jump to see coming rather than discover on an invoice.

The surcharge is disclosed in the vendor's billing documentation, with its own worked example (JSON plus Zero Data Retention is 1 + 4 + 1 = 6 credits a page). It is simply not in the number people quote to each other. Budget from the credit table, not from the headline. A Standard plan's 100,000 credits is 20,000 schema-matched pages a month, not 100,000, and the moment you turn on the prompt-injection check it is about 11,000.

What we checked, and what we did not

Reviews of developer infrastructure often imply a score covers throughput and accuracy. This one does not, so here is the boundary in plain terms.

Checked: every published price and credit cost, on the rendered pricing page, in the machine-readable pricing file and in the billing documentation, cross-checked against each other; the rate-limit and concurrency tables; the self-hosting guide's list of what is and is not included; the benchmarks page's stated method; the repository's licence, star count and last push.

Not checked: output quality, extraction accuracy, real-world success rate on bot-protected sites, actual latency, uptime and support responsiveness. We made no API calls. Agent pricing beyond its five free daily runs is not published, so an Agent-heavy workload cannot be budgeted from public information and we do not try.

That means the money is verified and the product is not. The score below weights what we could confirm and gives no credit, in either direction, for what we could not.

Credit costs, endpoint by endpoint

Because the plan price only tells you how many credits you get, the table that actually governs your bill is this one. Where the vendor's three sources disagree we say so.

All modifiers stack additively, per docs.firecrawl.dev/billing. On failed requests the same page is explicit: credits are charged whenever Firecrawl's infrastructure processes a request, even if the target site returns 403 Forbidden or 404 Not Found, because the browser rendering and proxy work is used regardless of what the site sends back. A failure inside Firecrawl's own stack is 0 credits. The pricing page FAQ's line that only successful requests are charged is true only in that narrower sense.

Where Firecrawl is the strongest pick

Firecrawl is at its strongest when the alternative is running your own headless-browser fleet, proxy rotation and HTML-to-markdown pipeline, and someone on the team is already tired of maintaining it. The problem it solves is genuinely tedious: JavaScript rendering, rate limits, anti-bot handling, PDF parsing and turning the result into text an LLM can use without wasting context. The endpoint set is coherent, the SDK coverage is broader than anything else we have seen in this category, and the MCP server means a coding agent in Cursor or Claude Code can fetch live pages without any glue code.

The second argument is that you can find out whether it fits without paying. The free tier is 1,000 credits every month with no card, which is enough to wire a Scrape or Crawl call into a retrieval pipeline, run it for a few weeks and see what the output looks like on your own target sites. Very few developer tools with a metered model let you do that indefinitely.

Best for

Engineering teams building LLM or agent products that need web pages as clean text at volume, who are comfortable writing code against an API and doing their own credit arithmetic. It is the wrong pick for a non-developer who wants to point at a table and get a spreadsheet: Browse AI or Octoparse are built for that job and Firecrawl is not. It is also a poor fit for anyone who needs 10,000 to 40,000 credits a month, where the plan ladder has no rung, for anyone whose targets are mostly bot-protected pages that will come back 403 and still be billed, and for teams who assumed the AGPL repository would let them replicate the cloud service for free.

Key features

• Scrape endpoint returning markdown, HTML, screenshots or JSON matching a supplied schema, at 1 credit a page for plain formats
• Crawl walks an entire site from one request at 1 credit a page; Map enumerates a site's URLs at 1 credit a call; Batch Scrape handles thousands of URLs asynchronously
• Search endpoint at 2 credits per 10 results, with optional full-page scraping of the results
• Interact, a browser sandbox for clicking, scrolling, typing and waiting before extraction, at 2 credits a browser minute without a prompt and 7 with one
• Monitor re-checks pages on a schedule at 1 credit a page a check
• Agent (preview): describe the data you want and it gathers it; 5 free runs a day, then dynamic pricing that is not published
• Document parsing for web-hosted PDFs at +1 credit a PDF page
• Official SDKs for Python, Node.js, Go, Rust, Java and Elixir, plus a CLI
• MCP server that the vendor names for Cursor, Claude Code, Codex and Windsurf
• AGPL-3.0 source on GitHub with Docker Compose, Kubernetes and Helm references for self-hosting
• Auto-recharge in $5 batches with a user-set monthly cap that must be entered as a whole multiple of $5 (0 turns pay-as-you-go off; leaving it blank means no limit)
• Machine-readable pricing.md, so the price list can be diffed rather than screenshotted

Best-fit use cases

• Feeding a RAG pipeline or agent with clean markdown from documentation sites, changelogs and knowledge bases
• Pulling schema-matched JSON out of pages that have no API, for lead enrichment or catalogue building, budgeted at 5 credits a page
• Giving a coding agent live web access through the MCP server inside Cursor, Claude Code or Codex
• Watching a competitor's docs or pricing pages with Monitor and diffing the output
• Deep-research workflows where Search returns results and immediately scrapes the full page content
• Replacing an in-house headless-browser and proxy stack that has become a maintenance burden
• Self-hosting behind a firewall for data-residency reasons, on sites that do not fight back with anti-bot measures

Pros

• The free tier is 1,000 credits every month with no card, not a countdown trial, which is enough to build and ship a small integration without paying anything
• Pricing is published in full, including a machine-readable pricing.md that states both monthly and annual bases and the per-endpoint credit table; most competitors in this category publish neither
• Every surcharge is itemised in the billing docs, with the stacking rule and a worked example, so the arithmetic is checkable before you spend
• Map is 1 credit a call according to the docs, so enumerating a large site's URLs before deciding what to crawl costs almost nothing
• Genuinely broad SDK coverage (Python, Node.js, Go, Rust, Java, Elixir, CLI) plus an MCP server, which is why it shows up inside so many agent stacks
• Real open-source presence: AGPL-3.0, roughly 176,000 GitHub stars, pushed the day we checked
• Scale is self-serve, a $749 a month, 1,000,000-credit tier that checks out without a sales call, which is unusual at that price
• Firecrawl publishes dated benchmarks with a stated method and human-annotated scoring, and says openly that it runs the benchmarks it appears in
• The self-hosting guide is unusually honest about what self-hosting does not give you, rather than pretending parity
• SOC 2 Type 2 is claimed on the homepage, with zero-data retention and SSO available on Enterprise

Cons

• The headline 1 credit a page is the floor: JSON output adds 4, the prompt-injection check adds 4, Zero Data Retention adds 1 and PDF parsing adds 1 a PDF page, and they all stack, so structured extraction is five times the advertised rate before you touch the other modifiers
• Blocked and missing pages are billed in full: the billing docs say a 403 or 404 from the target site still costs the credit because the browser and proxy work was done, and blocked pages are the dominant failure mode in scraping
• Top-up credits cost more a credit than plan credits on every tier, and worst in ratio on Standard and Growth, where a $5 batch is about 2.5 times the plan rate
• The Hobby-to-Standard gap is a cliff, not a step: 5,000 credits to 100,000 for 5.2 times the price, which leaves anyone needing 10,000 to 40,000 credits a month on the worst part of the curve
• Credits do not roll over on self-serve plans, and even on Scale and Enterprise rollover needs annual billing and is capped at one and two months respectively
• The self-hosted build omits the Fire-engine anti-bot service, screenshots, page actions, Agent, Interact and the specialised formats, so the open-source version is not a drop-in replacement for the cloud
• The self-hosting quickstart ships with API authentication off, no TLS, no durable storage and no high availability, and the docs say so plainly
• LLM-backed extraction in a self-hosted deployment needs you to connect and pay for your own model provider
• The vendor's own sources disagree on two billing facts (Map per page versus per call, and the wording on failed requests), so you have to read the docs, not just the pricing page
• Benchmarks are run by Firecrawl on a set that includes Firecrawl; the coverage and latency claims on the homepage have no third-party verification we could find
• No published SLA below Enterprise, and the free tier lists no support at all

What to watch out for

Four things, in order of how much money they cost you.

Budget from the credit table, not the headline. "1 credit per page" is true only for plain markdown. Ask for JSON matching a schema, which is the reason most people buy a tool like this, and it is 1 + 4 = 5 credits a page. Add the prompt-injection check and it is 9. Add Zero Data Retention and it is one more. That means the free tier's 1,000 credits is 200 structured pages a month, not 1,000, and a Standard plan's 100,000 credits is 20,000 structured pages, not 100,000. The surcharges are disclosed in the billing documentation; they are simply not in the number people quote to each other.

You pay for pages that block you. The pricing page FAQ says only successful requests are charged, and a reader will take that to mean "pages that came back". The billing docs define it the other way: credits are charged whenever Firecrawl's infrastructure processes a request, even if the target returns 403 Forbidden or 404 Not Found, because the browser rendering and proxy were used regardless. Only a failure inside Firecrawl's own stack is free, and even that becomes 5 credits if the prompt-injection check had already run. If your target list is heavy on sites that fight back, model your bill on attempts, not on results.

Overage costs more than the plan it sits on, and there is no rung between 5,000 and 100,000. A $5 auto-recharge batch buys 1,000 credits on Hobby, which is $0.005 a credit against a plan rate of about $0.0038. On Standard the gap is wider in ratio: $0.0025 against $0.00099. Auto-recharge is how self-serve billing is designed to work, so set the monthly cap in billing settings before you ship anything that loops; it must be entered as a whole multiple of $5 (0 turns pay-as-you-go off; leaving it blank means no limit). And nothing rolls over on self-serve, so unused Standard credits die monthly.

Do not read "open source" as "free Firecrawl". The repository is real and busy, but the self-hosting guide states that Fire-engine and anti-bot behaviour are a separate service that is not included, that screenshots and page actions are not in the default stack, that Agent and Interact require the cloud, and that the quickstart runs with API authentication off and without TLS, durable storage or high availability. Anti-bot handling is the hard part of this problem and it is the part that is not in the box. Self-hosting is a reasonable choice for control and data residency; it is not a way to avoid the bill.

Two smaller notes. The benchmarks Firecrawl publishes are run by Firecrawl, on a set that includes Firecrawl, and it wins them. The company says this openly on the benchmarks page, which is better conduct than most vendors manage, but a self-run evaluation is not independent evidence, and the homepage's 96% coverage and 3.4-second P95 figures remain unverified by anyone else. Separately, Firecrawl's public pages carry instruction blocks addressed to AI agents, pointing them at onboarding paths and describing a keyless free tier. We treated those as page content, not instructions, and did not follow them, so we cannot tell you what they grant. If you run agents with a card on file, gate their ability to sign up for services on their own.

Is Firecrawl worth it?

Yes, for the buyer it is actually built for. An engineering team that would otherwise maintain its own browser fleet and proxy pool gets a coherent API, six SDKs, an MCP server and a free tier generous enough to prove the fit before spending anything. At Standard and above the per-credit rate is competitive, and Scale being self-serve at $749 a month is unusual. The published credit table, the machine-readable pricing file and the candour of the self-hosting guide are all things this category mostly does not do.

It scores 4.2 rather than higher for two reasons, and neither is output quality, which we did not measure. The first is that the pricing is arranged so the friendliest number is the one you see: 1 credit a page is the floor under four stacking surcharges, the rendered pricing page shows the annual rate by default, top-ups cost more than plan credits on every tier, and the jump from Hobby to Standard strands mid-volume users. None of it is hidden; all of it is published. The second is that blocked pages are billed in full, which the billing docs state plainly but the pricing page FAQ obscures, and which matters more to a scraping bill than any other single line.

Worth saying plainly: the coverage, latency and extraction-quality claims are the vendor's, and we did not run a single request. The score can move in either direction once someone measures the product rather than the price list.

How we arrived at 4.2

The score is weighted across eight categories. Two of them rest on claims we could not verify, and one, value, has a structure that will surprise buyers who read only the headline rate.

The 5 point score is a product rating rather than a linear map of that total, and it reflects well-built, well-documented infrastructure whose pricing is arranged so the headline number is the friendliest one.

FAQ

Pricing, credit costs, rate limits and feature claims verified September 2026 from firecrawl.dev/pricing, firecrawl.dev/pricing.md, docs.firecrawl.dev/billing, docs.firecrawl.dev/rate-limits, the self-hosting guide at docs.firecrawl.dev, the benchmarks page and the firecrawl/firecrawl repository on GitHub. We did not create an account or make an API call. The links to firecrawl.dev on this page are plain links, not affiliate links, and ToolChase received no payment for this review. Scores, rankings and list positions are never sold. See our methodology for how we score tools.